facebookresearch/vggt ? reverse-engineered prompt

Build me a Python-based tool that can create a 3D scene from a set of 2D images. I want to be able to give it one or more photos of an object or place, and have it generate the key 3D information. It should be based on the VGGT model from Facebook AI research.

The main function is to take the images and figure out the camera positions and settings for each shot, create depth maps, and generate a 3D point cloud of the scene. It should use the pre-trained model weights from Hugging Face to do the heavy lifting.

Please create a simple Gradio web interface for this, so I can just upload my images and see the result. Include a way to visualize the reconstructed 3D point cloud directly in the app. Also, I need an option to export the output, especially the camera poses, into the COLMAP format so I can use it with other Gaussian splatting libraries.